High-precision Real-time Lightweighte Slam Based On RGB-D Camera

Unmanned driving technology is a huge and complex systematic project,but its core is always inseparable from sensors,localization,and navigation technologies.GPS technology,as the current mainstream localization technology,cannot meet centimeter-level positioning and obstacle avoidance,and cannot achieve full coverage indoors and in cities.The localization technology based on computer vision complements this shortcoming of GPS well.Simultaneous localization and mapping(SLAM)aims to realize 3D scene reconstruction and camera pose estimation of imaging sensors.The current mainstream SLAM algorithms have high requirements for computing performance,and often require high-performance GPU acceleration to achieve high-precision and high-frame-rate real-time operation.Due to the limitation of size and power consumption,this is not conducive to the popularization of SLAM algorithm on Portable devices.Therefore,this project will study the SLAM algorithm and make improvements to it for Portable devices to improve accuracy while reducing the amount of calculation.The main research work is as follows:1)Compared with the current mainstream SLAM system.The overall framework that is most suitable for indoor mobile devices is designed.RGB-D is used as input,the direct method is used for pose estimation,the loop detection method is used to remove the drift phenomenon of the SLAM system,the sparsity of SLAM is used,and the beam adjustment method is used to optimize the overall situation.In-depth study of the mathematical principles of each part of the SLAM system,study the factors affecting the accuracy and speed of the SLAM system,find the key parameters,and conduct targeted optimization for mobile devices.2)Preprocessing of RGB-D images.Preprocessing makes the depth map smooth,some noise and larger depth values will be removed.The commonly used algorithm is bilateral filtering.However,this method requires a large amount of calculation and is not suitable for use in mobile devices.This topic improves the preprocessing,compares the effect of different image noise reduction algorithms on the depth map,and finally Reflect the influence of positioning accuracy and speed.In the end,the preprocessing time was reduced by 87%without reducing the accuracy.3)Keyframe selection method.When the difference between the new image and the latest key frame is too large,a new key frame must be created.The selection of key frames is very important to balance the accuracy and speed of the system.The method of key frame selection aims to determine the appropriate key frame density.If the speed of the robot is slow,the frames will be dense and fewer key frames should be extracted.If the speed of the robot is fast,the image is sparse and more key frames should be extracted.In addition,if too many key frames are used,the amount of calculation will be wasted.However,if too few key frames are used,the accuracy will be too low.This subject proposes a new method of key frame selection.Considering that the SLAM system is more sensitive to pure rotation motion,the rotation and translation components are added to the judgment mechanism of the key frame selection method.And test it to determine the best value.After experimental tests,it is proved that the method effectively balances the accuracy and speed of the system.4)In the BA optimization at the back end of the system.Outliers are very common,often caused by moving objects or changes in shadows.It is necessary to use the Tukey Biweight robust kernel function to suppress its influence before optimization.The robust kernel function is optimized for mobile devices and indoor scenes,and the accuracy is improved without affecting the speed.Finally,the improved system was simulated and experimented,and compared with the latest method in the same public benchmark data set.The experiment proved that the overall improvement has achieved a 13%speed increase and a 12%accuracy increase on the original system.